launch-your-agent-claude-code

name	launch-your-agent-claude-code
description	Build, deploy, and iterate Claude Managed Agents from idea to production using Claude Code
triggers	["build a claude managed agent","launch my agent to production","create a scheduled claude agent","interview me about my agent idea","deploy agent to anthropic console","grade and iterate my agent","set up a recurring agent workflow","help me build on claude managed agents"]

launch-your-agent-claude-code

Skill by ara.so — AI Agent Skills collection.

Overview

launch-your-agent is a Claude Code skill that takes you from an agent idea to a live Claude Managed Agent (CMA) in your Anthropic Console. It orchestrates the full lifecycle: interview → scope v0 → launch in your account → grade against success criteria → iterate → schedule (if recurring).

You walk away with:

• A live managed agent in your Console
• Complete build artifacts in my-agent/ (build sheet, API payloads, eval scaffold, launch script)
• A graded initial run
• Scheduled deployment (if your task recurs)
• A NEXT-DIRECTIONS.md with v1/v2 roadmap

Primary Language: HTML (documentation/templates), with JavaScript/TypeScript for agent logic and shell scripts for deployment.

Installation

git clone https://github.com/anthropics/launch-your-agent.git
cd launch-your-agent
claude

The skills in .claude/skills/ are auto-loaded when you run Claude Code inside this directory.

Prerequisites

• Claude Code installed and authenticated
• Anthropic API key for your own account (create at platform.claude.com → API keys)
• API key stored in local .env file (never committed):

# .env
ANTHROPIC_API_KEY=sk-ant-api03-...

Key Commands

Main Flow

/launch-your-agent

Starts the full 4-phase flow:

1. Interview – Asks about your use case, inputs, outputs, success criteria, scheduling needs
2. Stage & Launch – Generates agent config, creates environment, deploys to your Console
3. Grade & Iterate – Runs eval against your success criteria, suggests improvements
4. Automate – Sets up scheduled deployment if task is recurring

Status Check / Close-Out

/wrap-up

Regenerates overview page, recaps all primitives you built, suggests next 1-2 upgrades.

Configuration

CMA Primitives

Claude Managed Agents support:

• Agent definition: System prompt, tools, model (claude-3-5-sonnet-20241022 or claude-3-7-sonnet-20250219)
• Environments: Isolated runtime contexts (dev/staging/prod)
• Secrets: API keys, credentials (encrypted at rest)
• Scheduled deployments: Cron-based recurring runs
• Tool use: API calls, database queries, file operations

Limits (as of documentation):

• Max system prompt: ~200k tokens
• Tools: Up to 2048 per agent
• Secrets: String values only, accessed via tool parameters

Reference: cma-primitives.md in repo root.

Interview Mapping

The interview phase captures:

Question	Maps to CMA primitive
What problem does this solve?	Agent system prompt (goals section)
What inputs does it need?	Tool definitions + environment secrets
What's a successful output?	Eval criteria + success rubric
How often should it run?	Scheduled deployment config (cron)
What should it NOT do?	System prompt (constraints/guardrails)

See interview-to-config.md for full mapping.

Code Examples

Agent Configuration Shape

// Generated in my-agent/agent-config.json
{
  "name": "daily-standup-summarizer",
  "model": "claude-3-5-sonnet-20241022",
  "system_prompt": "You are a standup summarizer. Each morning, you...",
  "tools": [
    {
      "name": "fetch_github_prs",
      "description": "Fetch open PRs from team repos",
      "input_schema": {
        "type": "object",
        "properties": {
          "repo": { "type": "string" },
          "github_token": { "type": "string" }
        },
        "required": ["repo", "github_token"]
      }
    }
  ],
  "max_tokens": 4096
}

Launch Script

#!/bin/bash
# my-agent/launch.sh

set -e

source .env

# 1. Create agent
AGENT_ID=$(curl -s https://api.anthropic.com/v1/agents \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "content-type: application/json" \
  -d @agent-config.json | jq -r '.id')

echo "Agent created: $AGENT_ID"

# 2. Create environment
ENV_ID=$(curl -s https://api.anthropic.com/v1/environments \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -d "{\"name\": \"production\", \"agent_id\": \"$AGENT_ID\"}" | jq -r '.id')

echo "Environment created: $ENV_ID"

# 3. Set secrets
curl -s https://api.anthropic.com/v1/environments/$ENV_ID/secrets \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -d "{\"key\": \"GITHUB_TOKEN\", \"value\": \"$GITHUB_TOKEN\"}"

# 4. Deploy
curl -s https://api.anthropic.com/v1/agents/$AGENT_ID/deploy \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -d "{\"environment_id\": \"$ENV_ID\"}"

echo "Deployed to $ENV_ID"

Eval Scaffold

// my-agent/eval.js
const Anthropic = require('@anthropic-ai/sdk');

async function gradeRun(runId, successCriteria) {
  const client = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY,
  });

  // Fetch run output
  const run = await client.agents.runs.retrieve(runId);
  const output = run.output;

  // Grade against criteria
  const gradePrompt = `
Success criteria:
${successCriteria.join('\n')}

Agent output:
${output}

Did the agent meet all criteria? Respond with JSON: {"pass": true/false, "feedback": "..."}
  `;

  const gradeResponse = await client.messages.create({
    model: 'claude-3-5-sonnet-20241022',
    max_tokens: 1024,
    messages: [{ role: 'user', content: gradePrompt }],
  });

  return JSON.parse(gradeResponse.content[0].text);
}

module.exports = { gradeRun };

Scheduled Deployment

// my-agent/schedule-config.json
{
  "agent_id": "agt_abc123",
  "environment_id": "env_xyz789",
  "cron": "0 9 * * 1-5",
  "timezone": "America/Los_Angeles"
}

# Apply schedule
curl https://api.anthropic.com/v1/schedules \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -d @schedule-config.json

Common Patterns

Pattern: Internal Workflow Agent

Use case: Daily standup summary from GitHub + Slack

1. Interview answers:

   • Problem: "Summarize team activity each morning"
   • Inputs: GitHub API token, Slack webhook
   • Success: "Covers all PRs, mentions blockers, <500 words"
   • Schedule: "Every weekday at 9am PT"

2. Generated tools: fetch_github_prs, post_slack_message

3. System prompt includes: goals, tone (concise), constraints (no speculation)

4. Scheduled deployment with cron 0 9 * * 1-5

Pattern: Customer-Facing Agent

Use case: Support ticket triage

1. Interview answers:

   • Problem: "Categorize incoming tickets, suggest help articles"
   • Inputs: Zendesk API, knowledge base embeddings
   • Success: "95% category accuracy, links 2+ relevant articles"
   • Schedule: "Real-time via webhook"

2. Generated tools: search_kb, update_ticket_tags

3. System prompt includes: customer empathy guidelines, escalation rules

4. No cron schedule (webhook-triggered)

Pattern: Data Pipeline Agent

Use case: Weekly analytics rollup

1. Interview answers:

   • Problem: "Aggregate usage metrics, detect anomalies"
   • Inputs: Postgres connection, previous week's baseline
   • Success: "Flags any >20% change, generates CSV"
   • Schedule: "Mondays at 6am UTC"

2. Generated tools: query_db, write_csv, send_email

3. System prompt includes: statistical thresholds, alert format

4. Scheduled deployment with cron 0 6 * * 1

Troubleshooting

"API key not found"

Ensure .env file exists in repo root with:

ANTHROPIC_API_KEY=sk-ant-api03-...

Never commit .env — it's in .gitignore by default.

"Agent creation failed: invalid tool schema"

Check my-agent/agent-config.json:

• input_schema must be valid JSON Schema
• required fields must be present in properties
• Tool names must be lowercase with underscores

"Eval keeps failing"

Refine success criteria in my-agent/build-sheet.md:

• Make criteria measurable (e.g., "mentions 3+ PRs" not "comprehensive")
• Add edge cases to test set
• Iterate system prompt with /launch-your-agent (picks up where it left off)

"Schedule not triggering"

Verify:

• Cron syntax valid (use https://crontab.guru)
• Timezone correct (defaults to UTC)
• Environment has required secrets set
• Check Console → Agents → Runs for error logs

"Rate limit hit during launch"

CMA API respects standard Anthropic rate limits. If hitting during deploy:

• Add sleep 2 between curl commands in launch.sh
• Or batch secret creation into single call

File Structure After Launch

my-agent/
├── build-sheet.md          # Interview answers + scoping decisions
├── agent-config.json       # Agent definition (system prompt, tools, model)
├── environment-config.json # Environment + secrets references
├── schedule-config.json    # Cron schedule (if recurring)
├── launch.sh               # Resumable deploy script
├── eval.js                 # Grading logic
├── OVERVIEW.md             # Human-readable status page
└── NEXT-DIRECTIONS.md      # v1/v2 roadmap

.env                        # API keys (git-ignored)

API Reference

See official docs: https://platform.claude.com/docs/en/managed-agents/overview

Key endpoints:

• POST /v1/agents – Create agent
• POST /v1/environments – Create environment
• POST /v1/environments/{id}/secrets – Set secrets
• POST /v1/agents/{id}/deploy – Deploy to environment
• POST /v1/schedules – Create cron schedule
• GET /v1/agents/runs/{id} – Fetch run output

Additional Resources

• cma-primitives.md – Full inventory of CMA features and limits
• interview-to-config.md – Detailed interview → config mapping
• examples-bank.md – Sourced agent examples and proof points
• ui/ – Example overview page templates

License

Apache 2.0 (see LICENSE in repo root)